[Klaud-Cold]

Goal: Update SGLang image from lmsysorg/sglang:v0.5.22-cu130 to lmsysorg/sglang:nightly-dev-cu13-20261027-5884e4fe.
Baseline: 2026-10-22 · lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae
AgentX · TP8/EP1 · Mean latency · Sources: API 1, API 2, API 3, API 4

Concurrency	Total tok/s/GPU ↑	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
1	2,589.77	18.62	507.13	3.23
4	3,972.89	27.77	354.88	3.54
8	7,105.89	51.66	376.95	4.08
12	8,128.1	63.01	484.04	5.51
16	12,403.8	82.17	1,078.58	6.9

Eval	Score ↑	Samples
gsm8k/em_strict · c16	96.97%	1,319

Stop: Close this PR at any time to stop this candidate now; Klaud cancels its runs and deletes its branch.

中文

**目标：**将 SGLang 镜像从 lmsysorg/sglang:v0.5.22-cu130 更新至 lmsysorg/sglang:nightly-dev-cu13-20261027-5884e4fe。
**基线：**2026-10-22 · lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae
AgentX · TP8/EP1 · 平均延迟 · 来源： API 1, API 2, API 3, API 4；数值及异常说明见上表。
**停止：**随时关闭此 PR 即可立即停止该候选：Klaud 会取消其运行并删除其分支。

[github-actions]

Thanks for the contribution!

• Review: If this PR changes files owned by someone other than a repository admin or @SemiAnalysisAI/core, ask one eligible CODEOWNER to complete the latest PR_REVIEW_CHECKLIST.md before contacting a core maintainer on Slack. Follow the template exactly, including As a PR reviewer and CODEOWNER, I have reviewed this and have, so sign-off verification triggers.
• PR verification: Sweeps run only when the PR appends an inferencex-e2e/perf-changelog.yaml entry and carries exactly one primary label: full-sweep-fail-fast (strongly recommended; canary plus per-matrix fail-fast), full-sweep-enabled (canary; matrix jobs continue after a failure), or non-canary-full-sweep-enabled (no canary or fail-fast). The modifiers all-evals, evals-only, and agentx-fast require a primary label. On fork PRs, a maintainer applies the label. See sweep labels and reuse.
• After merging: PR authors must ensure all GitHub Actions jobs pass. Transient failures often pass on rerun; see how to rerun failed jobs.

中文

感谢你的贡献！

• **审阅：**如果 PR 修改的文件归属于仓库管理员及 @SemiAnalysisAI/core 之外的 CODEOWNER，请先联系一位有资格的 CODEOWNER 填写最新的 PR_REVIEW_CHECKLIST.md，再通过 Slack 联系核心维护者。必须严格遵循模板，并保留 As a PR reviewer and CODEOWNER, I have reviewed this and have，才能触发签核验证。
• **PR 验证：**只有当 PR 在 inferencex-e2e/perf-changelog.yaml 末尾追加了条目，并且恰好带有一个主标签时，才会运行扫描：full-sweep-fail-fast（强烈推荐；canary 加逐矩阵 fail-fast）、full-sweep-enabled（有 canary；矩阵任务在失败后继续运行）或 non-canary-full-sweep-enabled（无 canary，也无 fail-fast）。修饰标签 all-evals、evals-only 和 agentx-fast 必须与主标签一起使用。fork PR 的标签由维护者添加。参见扫描标签与复用。
• **合并后：**PR 作者必须确保所有 GitHub Actions 任务通过。临时性失败通常可以通过重新运行恢复；参见重新运行失败任务的说明。

[Klaud-Cold]

Initial attempt · Passed · Run 40472636959 / attempt 1 · 2026-10-27 09:54 UTC
lmsysorg/sglang:nightly-dev-cu13-20261027-5884e4fe@sha256:61b9083bdcb7cf3741366f487cb996c05ea33ad2ddddc820d10cbc1c0cfb0a86 · 7ca692b9afea · AgentX · TP8/EP1 · Mean latency
Change: Bump the master image from v0.5.22-cu130 (release/v0.5.22 at baa9befa) to nightly-dev-cu13-20261027-5884e4fe (main at 5884e4fe, digest-pinned, measured ref 874df24); recipe flags unchanged because every recipe flag, glm47 tool parser and glm45 reasoning parser exists with the same semantics in the nightly (server_args/arg_groups diff only adds options), while flashinfer 0.7.1 to 0.7.2, sglang-kernel 0.4.9 to 0.4.10 and sgl-deep-gemm 0.2.2 to 0.2.3.post2 change underneath; sources: sgl-project/sglang@v0.5.22...5884e4f and Docker Hub lmsysorg/sglang tags, nightly runtime behaviour on GLM-5.2 NVFP4 is unverified until this smoke run.

Concurrency	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
1	18.59 (-0.1%)	554.48 (+9.3%)	3.2 (-0.9%)

Eval	Score ↑	Samples
gsm8k/em_strict · c16	97.27% (+0.30 pp)	1,319 each

Next: Smoke c1 benchmark (271/271 requests, 0 errors) and c16 gsm8k eval passed on 874df24; append the perf-changelog entry and dispatch the full AgentX final sweep.

中文

初次尝试 · 已通过 · Run 40472636959 / attempt 1 · 2026-10-27 09:54 UTC
lmsysorg/sglang:nightly-dev-cu13-20261027-5884e4fe@sha256:61b9083bdcb7cf3741366f487cb996c05ea33ad2ddddc820d10cbc1c0cfb0a86 · 7ca692b9afea · AgentX · TP8/EP1 · 平均延迟
**变更：**将主镜像从 v0.5.22-cu130（release/v0.5.22，baa9befa）更新为 nightly-dev-cu13-20261027-5884e4fe（main 5884e4fe，已按 digest 固定，测量 ref 874df24）；recipe 参数不变，因为所有 recipe 参数、glm47 工具调用解析器和 glm45 推理解析器在 nightly 中语义相同（server_args/arg_groups 差异只新增选项），底层 flashinfer 0.7.1 to 0.7.2、sglang-kernel 0.4.9 to 0.4.10、sgl-deep-gemm 0.2.2 to 0.2.3.post2 有变化；来源：sgl-project/sglang@v0.5.22...5884e4f 与 Docker Hub lmsysorg/sglang 标签，nightly 在 GLM-5.2 NVFP4 上的运行表现需本次 smoke 运行验证。实测数值及异常说明见上表。
**下一步：**874df24a 上的 c1 smoke 基准测试（271/271 请求，0 错误）和 c16 gsm8k 评测均已通过；接下来追加 perf-changelog 条目并启动完整 AgentX 最终扫描。

[Klaud-Cold]

parked · Session 1 is waiting on 40472636959. Recovery reruns preempted runs and starts the next session once they end.

中文

parked · 第 1 个会话正在等待 40472636959。恢复流程会重跑被抢占的运行，并在运行结束后启动下一个会话。

[Klaud-Cold]

resumed · The session ended before finishing; session 2 continues it in a new run of its planner.

中文

resumed · 会话未完成即已结束；第 2 个会话将在其规划运行的新运行中继续。

[Klaud-Cold]

Final full sweep · Passed · Run 40484444761 / attempt 1 · 2026-10-27 11:33 UTC
lmsysorg/sglang:nightly-dev-cu13-20261027-5884e4fe@sha256:61b9083bdcb7cf3741366f487cb996c05ea33ad2ddddc820d10cbc1c0cfb0a86 · a33421558d14 · AgentX · TP8/EP1 · Mean latency
Change: Validate the complete updated-image family at this head.

Concurrency	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
1	18.54 (-0.4%)	554.21 (+9.3%)	3.23 (+0%)
4	27.75 (-0.1%)	391.48 (+10.3%)	3.6 (+1.7%)
8	51.38 (-0.5%)	401.9 (+6.6%)	4.05 (-0.7%)
12	63.04 (+0%)	516.1 (+6.6%)	5.6 (+1.6%)
16	81.07 (-1.3%)	1,074.45 (-0.4%)	6.93 (+0.4%)

Note: c12, c16: 1 request error.

Eval	Score ↑	Samples
gsm8k/em_strict · c16	97.12% (+0.15 pp)	1,319 each

Next: Mark ready for maintainer review.

中文

最终完整 sweep · 已通过 · Run 40484444761 / attempt 1 · 2026-10-27 11:33 UTC
lmsysorg/sglang:nightly-dev-cu13-20261027-5884e4fe@sha256:61b9083bdcb7cf3741366f487cb996c05ea33ad2ddddc820d10cbc1c0cfb0a86 · a33421558d14 · AgentX · TP8/EP1 · 平均延迟
**变更：**验证此提交更新镜像后的完整配置族。实测数值及异常说明见上表。
**下一步：**标记为就绪，等待维护者审查。

[Klaud-Cold]

parked · Session 2 is waiting on 40484444761. Recovery reruns preempted runs and starts the next session once they end.

中文

parked · 第 2 个会话正在等待 40484444761。恢复流程会重跑被抢占的运行，并在运行结束后启动下一个会话。

[github-actions]

View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=40484444761

View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=40484444761

[Klaud-Cold]

validated · Repairs: unknown · Runs: 40472636959, 40484372016, 40484444761
All owned runs ended. Full sweep verified; ready for review.

中文

validated · 修复次数：未知 · 运行：40472636959, 40484372016, 40484444761
所有自有运行均已结束。完整 sweep 已验证；已就绪，等待审查。

[claude]

Claude finished @Klaud-Cold's task in 12s —— View job

---

LGTM - no blocking issues found.

The image change is a public Docker Hub tag pinned by digest. The perf-changelog.yaml entry for glm5.2-fp4-b200-sglang-agentic-mtp is added at the end of the file. The final AgentX sweep passed.

[claude]

Code review found no issues

No high-confidence issues detected in this change.